At the CVPR (Computer Vision and Pattern Recognition) 2022 conference earlier this year, Google Research presented its work on RawNeRF with NeRF in the Dark: High Dynamic Range View Synthesis from Noisy Raw Images.
A neural network that can convert 2D photos into a 3D scene is known as a neural radiance field (or NeRF). To concurrently denoise and reconstruct the scene, Google has developed one dubbed RawNeRF for dark scenes originally shared in June (h/t TechCrunch ). It mixes photos captured from numerous distinct camera perspectives.
The noise removal process used by RawNeRF, which uses RAW photos, is quite effective. The noise in the first case below can be sufficiently reduced by using Google’s solution so that the highlighted sign can be read.
Google’s RawNeRF can restore far more accurate color and detail throughout the entire scene when compared to other NeRFs. The degree of realism is getting close to the TV trope where characters request that an image be upgraded and the computer produces an astonishingly (and improbably) high-resolution result.
Google, meanwhile, is eager to point out that RawNeRF is more than just a denoiser and that it can also change the camera position to examine the image from different perspectives as well as exposure, tonemap, and focus (specifically, render synthetic defocus with accurate bokeh effects).
While the paper , data , and code are also accessible, Google’s six-minute video gives a solid introduction of RawNeRF. There is no guarantee that this work will ever be used in a product because it is still in the research stage. In order to take several pictures, users would need to enable RAW capture, which uses more space. Google Pictures does, however, provide Cinematic photos to mimic any 3D motion that may have occurred before a still image was taken.
L: RawNeRF | R: Original
FTC: We employ income-generating auto affiliate connections. MORE ON GOOGLE RESEARCH More.
CVPR 0
